Gaussian Process-assisted Evolutionary Algorithm for Constrained Expensive Multi-Objective Optimization

Abstract

It is well known that the surrogate-ssisted evolutionary algorithm has unique advantages in solving the computationally intensive multi-objective black-box optimization problems (EMOPs). SAEAs approximate objective functions with a surrogate model to reduce the number of function evaluations during optimization. However, many optimization problems are time-consuming for evaluating objectives and have computationally inexpensive constraint functions. Therefore, with few function evaluations, it is not easy to find many feasible solutions with good convergence. We propose a novel approach called GP-CMOEA that uses the cheapness of constraints to generate enough feasible solutions to solve this problem, ensuring that all expensive objective function evaluations are performed only for feasible solutions. Our experimental results demonstrate that the performance of the proposed algorithm with a small number of function evaluations is competitive.